mikgroup/sigpy

Sigpy compatibility with BART's *.cfl files

mscipio opened this issue · 10 comments

mscipio commented

I have a script that allows me to decode vendor-specific raw data from the Siemens Biograph mMR into .cfl files (kspace and non-cartesian trajectory) that can be fed into BART for the reconstruction.
I was wondering if sigpy has an interface that can allow me to use the cfl python library (e.g. function cfl.readcfl() and cfl.writecfl()) as input and use the way more pythonic tools provided by sigpy for the reconstruction.

sidward commented

You can find the cfl input and output files in the BART repo here: https://github.com/mrirecon/bart/tree/master/python

  • 👍1
frankong commented

Agree with @sidward that the BART repo already provides the functions, so it'll be best to use those instead of copying them to sigpy.

mscipio commented

Sorry to reopen this after such a long time, but I am still having issues here and there with decoding the proprietary files I can export from my Siemens scanner. Lately I have been trying to use the siemens_to_ismrmrd package to convert them to the ismrmrd format (from the native twix binary format in which I get them).

I was wondering if you had any experience using ismrmrd files as input to sigpy.
It should be possible to convert them to numpy array I believe, but I am not sure if some further manipulation is required? Has anyone ever tried this?

Why does sigpy not offer any kind of support for input files other than numpy?

sidward commented

I'm sorry, I have no experience with ismrmrd. Maybe you can try BART's twixread function?

As for why there are no plans to support input files other than numpy:

  • It's just a lot of overhead to ensure compatibility and I do not currently have the time to take that on.
  • There are existing repositories that supposedly address what you want: https://github.com/ismrmrd/ismrmrd-python

If you do create a ismrmrd_to_python repository that works better than the above, please let me know and I will link it in the README 😃

jtamir commented
sidward commented

Hi @mscipio I noticed you posted the issue in the BART repo as well. Have you tried mapVBVD at all?

In MATLAB, the "unsorted" option returns the raw acquired data, which you can re-order as needed.

mscipio commented

Hi @sidward

Yes, I have had relatively good success using mapvbvd.

I wanted to use another way to decode those data because sigpy is producing some "ghosting" artifacts in the reconstruction and I am having a hard time understanding what's happening.

Here is an example:
image

If you look at the rightmost subplot, it looks like the ghost image is flipped wrt the center of the image.

Right now I am getting the trajectory (which is a golden angle stack-of-spirals) from the log file of the scanner (Siemens mMR) and it looks like this:
image

I have a hard time believing I am doing something wrong with the trajectory (at the end of the day it's just a txt log file) so that's why I was trying to use something other than mapvbvd to decode the twix file, to see if that's was the problem

sidward commented

Are you acquiring multiple slices? If so, data are often acquired in a slice-interleaved manner, so you would need to manually re-order it.

mscipio commented

Yes, as you can see from the images above, it is a 3D volume.

you would need to manually re-order it.

Correct, that's what I have been trying to do. pymapvbvd and twixtools (both based on Matlab's mapvbvd) supposedly have the capability to correctly sort the acquisitions in the right multidimensional shape, but I guess that's not probably the case, given the results I am getting ...

sidward commented

If that's the case, that's something you would need to bring up with the maintainers of the above projects.

On my end, I have always used the unsorted data and manually re-ordered it. For Siemens, MATLAB's mapVBVD returns the sample locations which I use to sort the data.

obj = mapVBVD(dat_file);
lin = obj{numel(obj)}.image.Lin - obj{numel(obj)}.image.centerLin;
par = obj{numel(obj)}.image.Par - obj{numel(obj)}.image.centerPar;
tbl = obj{numel(obj)}.image.unsorted();

If your acquisition is volumetric, tbl will have dimensions along the lines of (sx, nc, nadc), where sx is the readout size, nc is the number of coils and nadc is the number of readouts. You can loop through nadc to build your k-space array. Keep in mind the for volumetric acquisitions, looping through nadc is looping through ky and kz, and for multi-slice acquisitions, looping through nadc is looping through ky and z. Additionally, you will need to re-order the z dimension appropriately to account for the interleaved acquisition.

Hope this helps.